> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://poly-ai.ferndocs.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://poly-ai.ferndocs.com/_mcp/server.

# 3.2026

> March 2026 release notes.

The **March 2026** PolyAI Agent Studio release improves how teams review and debug conversations, with deeper visibility into source content and large-scale conversation analysis.

Expand the items below for details:

#### Sources in Conversation Review

![source-hub-file-preview](/_fern-files/_img/0dc318784d1637a147ca0a935fedbc27e262e72ebbd31e26b9023fad6dbab201.webp)

Each conversation turn now shows a **Sources** tag listing the Connected Knowledge files the agent retrieved. Click a source name to open a side panel showing the exact document content used to generate the response.

**What's new:**

* **Sources** tag appears beneath each turn, alongside **Matched topics**.
* Click any source name to open an inline preview panel.
* The panel shows the retrieved text chunks from the source document.
* An **Open in Knowledge** button links directly to the source in the Knowledge area.
* Toggle **Sources** in the **Diagnosis** dropdown to show or hide this layer.

**Why it matters:**

* See exactly what content the agent was working from on any given turn.
* Faster debugging when retrieved content is unexpected or incomplete.
* No need to leave Conversation Review to check the source document.

#### Smart Analyst deep sampling

![deep-sampling](/_fern-files/_img/f4bb7a76ad2cb24d7b17806a21af92c2332a371251bfe6b56a4bccbf6e580ee8.webp)

Smart Analyst now uses a sub-agent to analyze up to **500 conversations** for patterns – a 10× increase from the previous 50-conversation limit.

**What's new:**

* Analyze up to **500 conversations** per query (previously 50).
* Conversations are sampled either **randomly** or **based on a metric / PolyScore**.
* Ask questions like *"What do customers complain about?"* or *"How can we reduce handoffs where users ask to speak to a human?"*
* Get **directional insights on behavior frequency** where no metric is present.
* Deep sampling helps uncover patterns that smaller samples might miss.

**Example use cases:**

* Identify common complaint themes across a large sample
* Analyze handoff reasons with statistically meaningful data
* Discover edge cases and rare conversation patterns
* Validate hypotheses about customer behavior at scale